Deep Learning-Based Spatiotemporal Fusion Approach for Producing High-Resolution NDVI Time-Series Datasets

The availability of concurrently high spatiotemporal resolution remote sensing data is highly desirable as they represent a key element for effective monitoring in various environmental applications. However, due to the tradeoff between the spatial resolution and acquisition frequency of current satellites, such data are still lacking. Many studies have been undertaken trying to overcome these problems; however, a couple of long-standing limitations remain, including accommodating abrupt temporal changes, dealing with complex and heterogeneous landscapes, and integrating other satellite datasets as well. Accordingly, this paper proposes a deep learning spatiotemporal data fusion approach based on Very Deep Super-Resolution (VDSR) to fuse the NDVI retrievals from Sentinel-2 and Landsat 8 images. The performances of VDSR are analyzed in comparison with those of two other classical methods, the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM) and the flexible spatiotemporal data fusion (FSDAF) method. The results obtained indicate that VDSR outperforms other data fusion algorithms as it generated the least blurred images and the most accurate predictions of synthetic NDVI values, particularly in areas with heterogeneous landscapes and abrupt land-cover changes. The proposed algorithm has broad prospects to improve near-real-time agricultural monitoring purposes and derivation of crop status conditions in the field-scale.